SemCity: Semantic Scene Generation with Triplane Diffusion

The team led by Jumin Lee advances the domain of 3D diffusion models with their system, SemCity, aiming at semantic scene generation in real-world outdoor settings. Dealing with real-world scenarios offers a significantly greater challenge due to sensor limitations, which their innovative triplane representation aims to resolve. The system showcases application in tasks like scene inpainting and outpainting, as well as semantic scene refinement, illustrating a notable step towards more interactive and immersive 3D virtual environments. Intricate details and implications of this approach are shared on their project page.

• Develops a 3D diffusion model for semantic outdoor scene generation
• Introduces a triplane representation to address data disposition challenges
• Facilitates object manipulation within scenes and expansion of environments

SemCity may redefine the expectations and capabilities associated with environmental and architectural modeling, game development, and urban planning simulations. By introducing a method that can handle the intricacies of real-world outdoor environments, SemCity could well be carving a new path for 3D modeling and scene generation technologies.